May_EDFA_Digital

edfas.org 33 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 21 NO. 2 examples are given in Fig. 2 to illustrate the general trend of acquiring images for different SEM settings. Note that the recorded time is for the entire IC. The increase in imaging time in varying field of view for smaller frame size (20 µm) is explained by the fact that it takes a larger number of frames to acquire the entire IC while the larger frames sizes (500 µm) can cover the same area in fewer frames. Fromthis table, the infeasibility of acquiring high- quality images inpractice canbeunderstood. Toovercome this limitation, image analysis algorithms can be applied to segment and extract features from low quality, quickly acquired images. Several factors make this a challenging problem for image analysis algorithms: • Feature shapes: The features that need to be extracted are of varying shapes and sizes. The lack of universality among features along with design rules present in ICs produced by different foundries make it difficult to develop a generic model for all features. • Manufacturing defects: Some ICs include inherent defects frommanufacturing processes [8] that might prevent them from being RE’d. Moreover, if the ICs are used, there may be additional abnormalities in the circuit from processes like electromigration. [9] • Deprocessing: Even though current technology can deprocess an IC with a high level of accuracy, several imperfections can cause additional artifacts in an image. This can be due to either warping of the IC during deprocessing [11] where materials are removed from the layer being imaged along with materials from the next layer, or improper removal of material from the layer currently being imaged. [10] There is also a possibility of oxidation on the metal structures in the IC caused by prolonged exposure to the envi- ronment during imaging. • Imaging modality: The noise introduced by the imaging modality is not only limited to sensor noise. In the case of SEM images, the noise also depends on the material being imaged and the topography/struc- ture of the image. Non-metal structures cause more spreading artifacts than metals. [11] With each of these noise sources acting in various intensities over a limited imaging space on the IC, it is unrealistic to expect generic image processing and analysis methods to produce satisfactory results for RE. Therefore, advancements in image analysis-based REwill require development of application-specific methods. APPLICABILITY OF SEM AND IMAGE ANALYSIS ALGORITHMS IN RE Earlier RE techniques relied on the knowledge of an SME to process images. Those techniques predominantly consisted of morphological transforms [11] such as local- ized erosion and dilation along with filtering techniques, e.g., median filters to denoise and extract features from the image. However, these techniques still required the images tobeof exceptional quality alongwith the constant attention of the SME on decoding the functionality of the gates visually one at a time. This was acceptable for larger nodes, but severely limiting for present-day technologies that integrate billions of transistors in the same die space. Significant limitations of the basic morphological transforms are now leading researchers to apply super- vised machine learning algorithms for segmentation and feature extraction from SEM images. This type of algo- rithm [12] learns andmodels the various structures in the IC image. Due to the inherent ability to learn shape andnoise models, which the basicmorphological transformcannot consider, supervised machine learning algorithms typi- cally outperformthem. However, toperformsatisfactorily, Fig. 2 Examples of images acquired at various settings. [17]